P300 Speller Performance Predictor Based on RSVP Multi-feature

Abstract

Brain-computer interface (BCI) systems were developed so that people can control computers or machines through their brain activity without moving their limbs. The P300 speller is one of the BCI applications used most commonly, as is very simple and reliable and can achieve satisfactory performance. However, like other BCIs, the P300 speller still has room for improvements in terms of its practical use, for example, selecting the best compromise between spelling accuracy and information transfer rate (ITR; speed) so that the P300 speller can maintain high accuracy while increasing spelling speed. Therefore, seeking correlates of, and predicting, the P300 speller's performance is necessary to understand and improve the technique. In this work, we investigated the correlations between rapid serial visual presentation (RSVP) task features and the P300 speller's performance. Fifty-five subjects participated in the RSVP and conventional matrix P300 speller tasks and RSVP behavioral and electroencephalography (EEG) features were compared in the P300's speller performance. We found that several of the RSVP's event-related potential (ERP) and behavioral features were correlated with the P300 speller's offline binary classification accuracy. Using these features, we propose a simple multi-feature performance predictor (r = 0.53, p = 0.0001) that outperforms any single feature performance predictor, including that of the conventional RSVP T1% predictor (r = 0.28, p = 0.06). This result demonstrates that selective multi-features can predict BCI performance better than a single feature alone.

Keywords: BCI; ERP; P300 speller; RSVP; performance variation; prediction.

Figure 1

Rapid serial visual presentation (RSVP) and P300 speller application. (A) RSVP paradigm. All subjects were instructed to focus on a green character (target) in a white-character stream and enter the target character using the keyboard during a total of 40 trials. (B) The conventional 6 × 6 matrix speller was implemented in BCI2000 software.

Figure 2

Event-related potential (ERP) features in the RSVP: P300 amplitude, P300 latency, trial variation in the P300 amplitude, and trial variation in the P300 latency. The bold line indicates the trial averaged P300, and dashed lines indicate single P300 trials (the same P300 response was presented for visual representation in this figure). For each electrode, P300 amplitude is defined as the time average of neighbor 50 ms with the peak amplitude (green dot: peak, red line: neighbor 50 ms with the peak), and the corresponding peak time point is regarded as the P300 latency. In addition, trial variations in the P300 amplitude and latency are defined as the inter-trial standard deviation of amplitude and latency, respectively.

Figure 3

P300 speller offline area under the ROC curve (AUC). This indicates sorted P300 speller offline performance (single-trial-based binary classification) that was estimated by 10-fold cross-validation.

Figure 4

RSVP features of high and low performers. Panel (A) shows three representative waveforms (black bold line) from the top 10 and bottom 10 performers with their single-trial waveforms: thin dashed lines (for visual representation, 10 single-trial waveforms selected randomly were drawn). Panel (B) indicates ERP waveforms of all of the high and low performers. Panel (C) compares group-difference through the scalp topography maps representing the averaged spatial patterns of the RSVP electroencephalography (EEG) features for high and low performers, and the box plots show the statistical difference between high and low performers. The red asterisk denotes statistical significance as shown at the bottom. To compute spatial patterns (scalp topography) and group differences (box plot), single-electrode features and averaged electrode (Fz, Cz, Pz, CP1, and CP2) features were used, respectively.

Figure 5

Association with P300 speller performance. (A) Relationship between the various RSVP features and P300 speller performance using Pearson’s correlation coefficients and linear regression analysis. RSVP T1% is a behavioral feature and the others are RSVP EEG features. All except P300 latency and trial variation in P300 amplitude were correlated significantly with P300 speller performance. (B) The spatial patterns of the relationship between RSVP EEG features and P300 speller performance. Pearson’s correlation coefficients and their statistical significance (uncorrected p-values) are presented in the top and bottom row, respectively. The green dots in the scalp topography in the bottom row denote significant channels after FDR-correction.

Figure 6

Relation between multi-feature predictors and P300 speller performance. The relationship with P300 speller performance are displayed for the multi-feature regular model (A) and stepwise model (B), respectively.

Figure 7

The correlation changes over the number of trials. This represents the way correlations change when the number of RSVP trials used to calculate EEG features varies. The X-axis represents the number of trials (1st to Nth RSVP trial), and the Y-axis represents the Pearson’s correlation coefficient’s absolute value.

Figure 8

Data distribution of each RSVP feature. The distribution of each of the RSVP features is depicted using relative frequency histograms (N = 48). Compared to the other RSVP features, RSVP T1% showed a sharp and sparse distribution.